Claude Science is a new AI workbench built specifically for researchers

Scientific research has always been slow, partly because of the science itself and partly because of the tools. Researchers regularly jump between PubMed, Jupyter notebooks, R, cluster terminals, and dozens of specialist databases, each speaking its own data language. The actual thinking gets interrupted by housekeeping.

Anthropic wants to fix that. On June 30, 2026, the company launched Claude Science, an AI workbench that pulls those fragmented tools into a single research environment. It's available now in beta for Claude Pro, Max, Team, and Enterprise users on macOS and Linux.

The launch follows Anthropic's push into life sciences that started last fall, and it's the most significant expansion of those efforts so far. The company has been building model capabilities, connecting to the scientific ecosystem through MCPs and skills, and signing research partnerships. Claude Science is where that work comes together.

How does it work?

Claude Science runs as an app, either locally on your machine or on a remote server over SSH or an HPC login node. Scientists interact with a coordinating AI agent that has access to more than 60 pre-configured tools and connectors covering:

• Genomics and single-cell analysis
• Proteomics and structural biology
• Cheminformatics
• More than 60 scientific databases including UniProt, PDB, Ensembl, ClinVar, and ChEMBL

One of the more practical features is how it handles reproducibility. When Claude Science generates a figure, it packages the exact code that created it, the computing environment, a plain-language explanation of how it was built, and the full message history. You can ask it to change an axis to log scale or remove gridlines in plain language, and the agent edits its own code to match. Everything stays traceable.

Compute management is also built in. For large jobs like protein folding or running a genomics pipeline over a big dataset, the app drafts a plan, asks for your approval before consuming new resources, then writes and submits the job to your existing HPC cluster over SSH or to Modal for on-demand compute. It can scale from one GPU to hundreds. Because agents run inside a persistent session, large datasets only need to load once.

A separate reviewer agent runs in the background throughout, checking citations, flagging numbers that can't be traced, and catching figures that don't match their underlying code. It self-corrects as it goes.

The app also connects natively to NVIDIA's BioNeMo Agent Toolkit, which gives it access to life sciences models including Evo 2, Boltz-2, and OpenFold3. Labs can also bring their own pipelines and save them as reusable skills for future sessions.

Why does it matter?

The early results from beta users suggest the productivity gains are real. A few examples:

• Manifold Bio, which designs tissue-targeting medicines, used Claude Science to run end-to-end target assessment, evaluating surface expression, trafficking, and safety data for hundreds of targets at once, drawing on the company's own internal data.
• Jerome Lecoq, a neuroscientist at the Allen Institute, built a multi-agent pipeline using Claude Science to write long-form scientific reviews. Sub-agents read thousands of papers, extract key claims and quantitative findings, build an evidence database, then draft the review section by section with dedicated agents generating cross-study figures. Reviews that previously took his team up to two years now take a fraction of the time. He has about 10 reviews completed, many over 100 pages.
• Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, used Claude Science for research on the molecular epidemiology of glioma. He says the app reduced the time needed for comprehensive germline analysis to roughly one-tenth of what it previously took, and his team independently validated the results.

Reproducibility and auditability are the key things that separate Claude Science from a general coding assistant. Science depends on other researchers being able to check and replicate your work. By baking that traceability into every output, Anthropic is targeting one of the genuine friction points in computational research rather than just making it faster to write code.

The context

AI tools aimed at scientific research have been a crowded space for a while, with companies like Google DeepMind, Microsoft, and a wave of biotech startups all competing to accelerate drug discovery and genomics research. What's different here is the scope. Claude Science isn't built around a single model or a single task. It's an attempt to replace the entire fragmented toolkit that researchers currently piece together themselves.

Anthropic is also putting money behind the launch. The company is supporting up to 50 research projects through its Claude Science AI for Science program, offering up to $30,000 in credits per project. Modal is adding up to $2,000 in compute for selected projects. Applications are open through July 15, 2026, with awards announced by July 31. Projects run from September 1 to December 1, 2026. The early focus is on biology and biomedical research. You can apply at claude.com/science.

There's also a discounted Team plan for academic institutions and nonprofit research organizations, which brings the pricing closer to what research labs can actually afford. That's a meaningful detail: powerful AI tools that only large pharmaceutical companies can pay for don't do much for the broader pace of scientific discovery.